Everything Point has learned about how you write sits in one list you can open. It reads in about a minute. You can change any line, and you can take out anything that’s stopped being true. This page is about that minute: what to look at, how to tell what went wrong, and what changes when you delete a line.
One idea carries most of the weight. A line in that list is an instruction, not a description of you. It runs on messages you haven’t seen yet. The two read the same on the page, and they fail in completely different ways.
- Test a line by what it does to the next message it fires on. An accurate sentence can still be a bad rule.
- Most of a good review is removal. Adding is the part people expect, and it’s the smaller half.
- When a draft comes back wrong, three things could be behind it, and one of the three is fixed in the list. Storing a one-off repair as a standing rule is the commonest way a list goes bad.
- Delete a rule and that decision goes back to what your own sent mail already shows, which is usually close.
- Plenty of what shapes your inbox lives elsewhere. Who counts as important, how much Point handles on its own, and what sits at the top of your feed are set on other screens.
What one line in that list actually does
What you open is a set of plain sentences, newest at the top. The ones you typed yourself are marked as added by you. They’re in your own words, and you could read any of them out loud to somebody.
How they got there is how Point learns your habits. You write one. You say something in passing and Point offers it back. You correct a piece of work and the correction is kept. This page picks up after that, with things already in the list.
Each line is an instruction with a trigger. It waits until a matching message arrives, then it fires. What’s inside the sentence decides how wide it reaches: one correspondent, one kind of message, or everything you write. A line reading keep it short and plain with Marcus stays quiet on every thread except one. A line reading keep it short and plain runs on your letters to the bank, your fee notices, and the note you send after somebody’s father dies.
That’s the difference between a list you can read and a list you’ve read. Scanning it and thinking yes, that sounds like me is the easy pass. Reading it means asking what each line does.
There’s a practical consequence worth having in advance. When a draft comes back in a shape you wanted otherwise, you find the line responsible by reading the list against the draft and asking which of these sentences could have produced this. That takes a few seconds on a list of eight lines. On a list of forty it’s real work. That’s the strongest everyday argument for keeping it short.
What changes when you can see the rules
Whether a visible, editable list is worth anything has been measured, on something close enough to email to be useful.
Todd Kulesza, Margaret Burnett, Weng-Keen Wong and Simone Stumpf presented Principles of Explanatory Debugging to Personalize Interactive Machine Learning at the 2015 conference on Intelligent User Interfaces. They gave 77 people 30 minutes with a message-sorting prototype and asked them to make its predictions as accurate as possible. Half of them worked with a version that showed its reasoning and let them correct it directly. The other half got the same tool with the explanations and the direct corrections stripped out, so their only move was to keep labeling messages and hope.
The people who could see the reasoning scored 52 percent higher on a test of how well they understood the system. Each correction they made was worth about twice as much as a correction from the other group. They also did less work to get there. They labeled an average of 47 messages against 182, and opened 151 messages against 296. Their classifiers ended up more accurate, at 0.85 against 0.77 on the paper’s combined measure of precision and recall.
The messages were public newsgroup posts about baseball and hockey rather than your mail, and the participants were adjusting word weights rather than sentences they’d written. So the numbers stay with that study. What carries over is the shape of the result. When people could see what the system was working from, the same amount of effort bought roughly twice the correction, and they stopped guessing. A participant from that second group summed the condition up, describing the software as “annoying” but saying that working with it “would have been easier if we knew how it made predictions.”
That’s the case for spending the minute. Sight makes correction cheap.
True about you is not the same as usable
Here’s the failure that belongs to reviewing a list rather than writing one. You read a line, you recognize yourself in it, and you leave it alone. It goes on quietly doing more than you meant.
I keep client email short is true of you. As an instruction it applies at every length, on every occasion, to nearly everyone you write to, since a great deal of your mail is arguably client mail. Three properties make a rule usable. Truth is a fourth, and it’s the one that fools you.
Who it fires on. A name, a company, or a category. Name one, or it fires on everyone.
When it fires. An occasion narrows a rule as well as an audience does. When I’m answering a question about fees is a fence. Always is open ground.
Whether it has one reading only. Open with their first name has exactly one. Be warmer has as many readings as there are drafts, and you’ll spend two weeks finding out which one it settled on.
One check catches all three, and it takes ten seconds a line. Read the sentence as though you’re handing it to a competent new assistant on their first morning, someone who knows your practice only through that sentence. Ask what they’d do with your next twenty messages. If you can predict that, so can the software. If you’re guessing, what you’ve written is a mood.
Then check for collisions, which start once there’s more than one line. Two rules can fire on the same message and pull in different directions, and that’s usually fine: the narrower one ought to win. The pair to look for is the one where you’d struggle to say which should win. That pair is a decision still waiting on you, and the fix is to merge them into one sentence that states the answer.
Writing a good rule in the first place, with the audience inside the sentence rather than in your head, is worked through in getting AI replies that sound like you wrote them and in when a summary learns from the corrections you make. The job here is harder. You’re judging a sentence already in force, written by a version of you who had one particular message on the screen and is out of reach now.
Three things that look identical from a bad draft
A draft comes back wrong. It’s a draft, so you have time. Work out which of three situations you’re in first, because the right response to each one makes the other two worse.
A rule did exactly what it says, and it went further than you meant. The tell is consistency. The wrongness turns up on every message of a certain kind, it points at one of your lines, and it reads as more of that line rather than a departure from it. Narrow the sentence that caused it. A second sentence pulling the other way leaves you with two live instructions and a draft that surprises you again later.
A rule is right, and this message sat outside it. The tell is that the message falls outside the scope you named. A per-person rule stays quiet on a stranger’s thread, and a rule about fee questions stays quiet on a scheduling note. Everything here is working. If you meant the rule to cover this case, widen the existing sentence deliberately. Writing the same thing again in a new line gives you two rules that agree today and disagree the first time you edit one.
A fact came out wrong, and the rules are beside the point. A date lifted from the sixth message in a thread, where it was a proposal rather than a settlement. A figure carried over from last year’s engagement, accurate and stale at once. The tell is that you can read every line in your list and see that each one is silent on this, which is why reading the list before acting is worth the few seconds it costs.
That third case is the one people handle badly, and it’s worth being blunt about why. The instinct after being burned by a draft is to write a rule so it never happens again. A rule shapes how something is written, and a date is either right or wrong, so the sentence you’d write sits there forever and clutters every review you do from here on. What you found is a mistake rather than a habit, and mistakes get corrected in the draft. The boundary between what a learned habit reaches and what sits outside it is set out in how Point learns your habits. The short form: register and accuracy are separate things, and tuning the first leaves the second exactly where it was.
Why editing a line beats adding another
When a rule is nearly right, change the words in it. That sounds obvious, and it’s the single most common thing people get wrong. Adding feels safe. Editing feels like undoing your own work.
Adding carries its own cost. Two sentences that overlap will eventually both fire on one message, and then you’re relying on a tie-break you never specified. Come back in October to review the list and the two lines look like two considered decisions rather than one decision and one patch. You keep both.
So: one idea per line. If the edit you’re about to make needs the word and to hold it together, you have two rules, and they should be written as two, each with its own scope. And if you’re writing a third sentence on a subject that already has two, stop. Three attempts at the same instruction is a sign that what you want lives outside a standing rule, and the right move is to delete all three and handle those messages yourself.
The other half of editing takes people by surprise. In the Kulesza study above, the group that could see what the system was using did add plenty, averaging 34.5 new terms. They also removed 8.2, and 7.4 of those were among the ten the system had started with. Given sight of what a system was actually working from, people took out most of its original contents. Carry that to your own list as an expectation rather than as a target: a review that keeps every line is usually a review that skipped the reading.
Deletion is the default move for any line you can justify only by reconstructing the day you wrote it. If you have to rebuild the situation that produced a rule before you can decide whether to keep it, you have your answer. Rules are cheap to make again, and the list holds the small number of things that stay true and that your mail alone would miss, which is why a short list is the sign it is working. A short list is where a working one ends up.
What deleting a line takes back
Three things are worth knowing before you start cutting, and all three make cutting easier than it looks.
It works forward, not backward. Removing a rule changes the work prepared from then on. A draft already written stays as it is, and a message already sent stays where it went. That last point is a real limit rather than a policy: a reply sitting on somebody else’s server belongs to that server now, which is where undoing what Point did draws its line.
Your own habits are underneath it. This is the part that stops people. A rule in the list is an exception to what your own correspondence already shows. Take it out and that decision goes back to the inference from years of your own sent mail, which is usually close and occasionally better than the rule was. You’re reverting to yourself.
Your mail and your record stay put. A rule only ever shaped how something was written or read, so the messages themselves are exactly what was said. The account of what was done on your behalf is a separate record with a separate job, covered in the activity log behind every action.
One outcome is worth watching for. Delete a line, and if the next week’s drafts slip in a way you’d have struggled to predict, that line was doing more than its wording suggested, and now you know it. Write the replacement narrowly, describing the effect you actually miss. The original sentence was evidently vague enough to be doing two jobs.
The rules that were right in March
A rule stays in force until you take it out, and a stale one keeps producing drafts that read fine. That’s exactly the problem: fine drafts are what keep it alive.
Most preferences are made on an occasion, and the occasion ends first.
- A ceiling on length written in the middle of busy season, still in force in July when you have time to write properly.
- A register set for a client who has since moved to a new firm, or for a contact who has left the company and taken the relationship with them.
- A phrase you banned in a bad week and have quietly gone back to using.
- Anything written for a project, an engagement or a business you’ve finished with.
Which means the review happens on a date rather than on a feeling. There’s a good field study on what a date is worth, and it’s about phones rather than email, which is worth saying up front.
Hazim Almuhimedi, Florian Schaub, Norman Sadeh, Idris Adjerid, Alessandro Acquisti, Joshua Gluck, Lorrie Faith Cranor and Yuvraj Agarwal presented Your Location has been Shared 5,398 Times! at the 2015 conference on Human Factors in Computing Systems. In it, 23 people ran a 22-day study on their own Android phones. For one week they had an app permission manager available and that alone. They used it. By Carnegie Mellon’s own account of the study, participants collectively reviewed their permissions 51 times and restricted 272 of them across 76 apps. Then they stopped. “Once the participants had set their preferences over the first few days, they stopped making changes.”
In the last eight days they were sent a daily message telling them how often their data had actually been accessed. Reviewing restarted, with 69 further visits and 122 more permissions blocked across 47 apps. In the authors’ own summary, 95 percent of participants reassessed their permissions and 58 percent restricted some further.
Those are privacy settings on a phone, and the stakes sit somewhere else from preferences about how your email gets written. What carries over is the behavior. People open a panel once and then leave it alone. What restarted the reviewing was the sight of what the standing decisions had been doing all along.
The prompt to look is yours to set, so put it on a date you already keep. The first of the month, or the morning you send invoices. Two minutes, and one question per line: did I write this for a situation that has ended?
What this list does not govern
A fair amount of what shapes your inbox is set somewhere other than this list, and going to the memory list to change it costs you an afternoon. Four things live elsewhere.
Who matters. Marking somebody important is its own act, with its own consequences for what you see first, and it lives on its own screen. Day four: marking the VIPs in your email is where that gets done.
How much Point handles on its own. The autonomy dial is held separately for each type of work and runs from suggest-only, through review, to fully handled, with drafting starting at review. The dial moves when you move it, and the memory list leaves it exactly where it sits. That decision lives in setting how much your inbox does on its own, and the case for leaving drafting where it starts is why a draft waits for you by default.
What’s true. Rules change how something is written or read. The messages underneath stay exactly what was said, and a fact that came out wrong in a draft is corrected in the draft.
Where your mail goes. Whether text you write is used to fit a model is a separate arrangement from whether a sentence is stored against your account, and it’s the one you’d have to answer to a client about. Keeping your data out of AI training is the page for that, and it’s worth settling before the rest of this matters.
The blunt version: if what’s bothering you is that the wrong things are at the top of your inbox, that order is decided on a different screen.
How Point shows what it holds
Point is an email client that sits on the Gmail or Microsoft 365 mailbox you use today. Same address, same history, everything where you left it.
Point’s whole record of your preferences is one panel of ordinary sentences, newest at the top, with the ones you typed marked as added by you. Every line can be edited and every line can be removed, at any point, in a click. Adding one takes a single plain sentence, saved and applied from then on. That panel is the whole of it, written in your own language, which is the condition that makes a monthly review a real check rather than a gesture.
Rules travel with the work, wherever Point is doing something for you. That’s what makes a one-line change worth making, and it’s also why scope deserves the ten seconds. One sentence about how you sign off to clients shows up under a reply on a thread where the subject never came up. Say in passing that a correspondent is a friend rather than a client, and Point offers that back as a proposed rule in plain words. A click later, the next reply to that person is written in the register you named.
Drafting starts at review. The words are prepared and the sending waits for you, until you move that setting yourself, one kind of work at a time. What was done on your behalf is written down in ordinary words as it happens, and a reversal reaches whatever is still yours to reverse, which a message already delivered elsewhere is out of. The benefits page covers the rest of what’s finished by the time you sit down, and the argument for the remembering living inside the client rather than in a separate tab is what an AI email client is.
Common questions
Where do I see what it has learned about me, and can I change it?
In one list, in plain sentences, newest first, with the lines you added yourself marked as yours. Each one can be edited or removed whenever you want, and adding one takes a single sentence. The reason it matters more than it sounds is diagnostic rather than philosophical. Reading the list is how you tell a rule that reached further than you meant from a plain mistake about a fact, and those two call for opposite responses.
A draft came out wrong. How do I tell whether the list caused it?
Read the list against the draft and ask which of these sentences could have produced this. If one of them could, and the wrongness holds across messages of that kind, narrow that sentence. If the list comes up clean, you’re looking at a mistake about a fact rather than a preference, and it gets corrected in the draft itself. That check is quick on a short list and heavy going on a long one, which is most of the case for pruning.
Should I edit an existing line or write a new one?
Edit, almost always. A second sentence next to a first leaves two live instructions that will eventually fire on the same message and disagree, and by the time that happens the patch looks like a decision. Add a new line when the subject is genuinely new. Edit when you’re adjusting something already covered. And if you’re on your third attempt at the same instruction, the thing you want probably lives outside a standing rule.
What happens if I delete a rule, or delete all of them?
Deleting works forward. It changes what’s prepared afterward, and it leaves anything already written or sent exactly as it stands. Each line was an exception to what your own sent mail already shows, so removing it hands that decision back to years of your own correspondence. An empty list is a working state. Drafts still come back in roughly your shape, and when one is wrong the same way every time, that’s the signal a sentence is worth writing again.
There is a line in there I do not remember writing.
That’s normal, and it’s a sign of things going right. Most rules arrive from something you said in passing while looking at a piece of work. Point offers it back, you keep it with a click, and a click is a much lighter act than opening a settings screen, so it’s easy to forget. Treat it the same as any other line. Read it as an instruction, work out what it does to your next twenty messages, and delete it if you can justify it only by reconstructing the day.
How often is this worth reading?
A couple of minutes a month, attached to something you already do, is enough for most mailboxes. The reason to schedule it rather than wait until something annoys you is that a stale rule produces perfectly reasonable drafts. They’re written to a situation that ended in March, they read fine, and so they stay.
Does the list show everything the system has learned?
It holds the preferences and habits, in sentences. The ordering of your inbox, who counts as important, and how much gets done for you are each decided on their own screen, and knowing that saves you a trip. It’s worth knowing the boundary before you conclude the list is wrong about you: quite often the thing you want to change is real, and it’s waiting on a different screen.
The short version
The list is instructions. Read each line by asking what it does to the next message it fires on, rather than whether it sounds like you. A good rule names an audience, an occasion, and one way to be satisfied. A sentence can be entirely true about you and still be a bad rule when it carries none of the three. Hand it to an imaginary new assistant and see if you can predict their week.
When a draft is wrong, sort it before you act. A rule that overreached gets narrowed. A rule that sat outside the message gets widened, or left where it is. A wrong date is a mistake rather than a preference, and it belongs far away from the list, which is where most lists start going bad. Edit rather than add, expect to remove more than you write, and delete anything you have to reconstruct a reason for. What you take out leaves your own sent mail underneath, still there.
Then put the review on a date, because a stale rule keeps working quietly for a situation that has ended and hands you fine drafts the whole time. What you end up keeping depends on the desk. An accounting practice tends to hold rules about brevity and about what has to appear above the fold. A consultancy tends to hold rules about how hard a sentence is allowed to commit. If your desk is a third thing, there are other versions of the same morning.